An AI count of dividing cells helped three pathologists agree more consistently on the grade of urothelial cancers in a multi-institutional study. The improvement appeared in both internal and external cases, giving the research a concrete clinical contribution: a common measurement reduced variation between expert readers.[1]
Researchers developed a deep-learning detector using 91,183 image patches associated with 318,893 annotations, then evaluated its use with three pathologists across two cohorts. The internal cohort contained 149 cases and the external cohort 316, for 465 in total. Each reader examined cases in an unassisted and an AI-supported phase. That design matters: it assesses not just a standalone algorithm's performance but what happens when the algorithm is placed alongside the specialist.[1]
A shared measurement for a difficult specimen
Urothelial carcinoma arises in the lining of the urinary tract, commonly the bladder. Histological grading is one part of assessing the tumour and planning care. Mitoses—cells visibly dividing—can offer an objective-looking signal, but the tissue landscape remains varied. Some dense or damaged areas can make mitotic figures difficult to identify. Other architectural and nuclear features remain relevant to an overall grade. If a model highlights additional dividing cells, the pathologist still has to judge their significance in context.[1]
The AI-derived mitotic count reached an AUC of 0.91 for distinguishing high-grade carcinoma across the combined cohort. The study used a threshold of 17.5 mitoses per two square millimetres; that is the boundary evaluated in this experiment, not a replacement for the wider tissue assessment used in grading.[1]
With AI assistance, mean reader AUC increased from 0.90 to 0.98 internally. The external increase was from 0.84 to 0.95, but its p-value of 0.11 did not meet the usual statistical-significance threshold. The agreement result was clearer: Fleiss’s kappa improved significantly in both cohorts.[1]
What changed with AI assistance
| Measure | Internal cohort: 149 cases | External cohort: 316 cases |
|---|---|---|
| Mean reader AUC | 0.90 → 0.98; p = 0.038 | 0.84 → 0.95; p = 0.11 |
| Fleiss’s kappa | 0.72 → 0.93 | 0.62 → 0.85 |
Agreement may matter as much as a better score
Fleiss’s kappa measures agreement beyond chance. Its improvement indicates that readers’ grades became more consistent when they had the quantitative aid. That is distinct from AUC, which evaluates discrimination between high- and lower-grade cases; the study is stronger when the two results are reported separately.[1]
That result has a practical logic. If three specialists count different sets of dividing cells, their grades may diverge partly because their attention was directed to different microscopic regions. A tool that consistently marks candidate mitoses can give them a shared reference. Agreement is not the same as correctness, however. Three readers can become more consistent around the wrong answer if the assistance systematically misses a tumour subtype or focuses on irrelevant tissue. A reliable evaluation must track both agreement and the independent diagnostic standard.[1]
The next step is routine slide-reading
The detector was trained on image patches but is intended to quantify mitoses across whole slides. External validation helps test that transition beyond the development collection. Prospective use would show whether the gain in reader agreement persists across routine scanning, staining and tissue variation.[1]
The promising role is as a shared measuring instrument. By giving pathologists a consistent view of mitotic activity, the system can reduce one source of disagreement while leaving the final interpretation of the specimen with the specialist.[1]
Sources & further reading
- Deep learning-based mitotic quantification for objective adjunct for grading of urothelial carcinoma: a multi-institutional validation study Scientific Reports · 22 Oct 2026 · Accessed 26 Oct 2026